Autonomy: The Quest to Build the Driverless Car - And How It Will Reshape Our World. Lawrence Burns
Читать онлайн книгу.Autonomy: The Quest to Build the Driverless Car - And How It Will Reshape Our World - Lawrence Burns
the hardware. A synchronization board regulated the timing of the information coming in from each disparate sensor, allowing Boss’s computing cluster to build a simulation of 3-D reality, the same way human drivers use their eyes and ears to build a model of the world in their heads. Another step involved predicting the behavior of other objects. For cars to be truly autonomous, they would have to be able to anticipate the behavior of practitioners of many different modes of urban transportation, from pedestrians to cyclists, skateboarders and scooter riders, among many others. But DARPA had told its teams that this race would happen in a vastly simplified environment; the only moving objects on the suburban course would be other automobiles. That made things a lot easier for the men and women coding the software, because it meant that Boss had to understand only a single set of behaviors—a tendency to move only forward and back along curvilinear lines, for example. Everything else, to Boss, was a stationary object.
As 2006 gave way to 2007, I flew to Pittsburgh to visit Red, Urmson, Salesky and the rest of the Tartan Racing team members at Robot City, including the GM engineers working with the team. I can remember touring the team’s workspace in the winter, amid these post-industrial ruins, the railroad roundhouse, the trailer set on the frozen ground, and being struck by how chilly the offices were. Everybody walked around in winter coats with knit caps on their heads. You could see your breath. These guys are really living lean, I thought. Rather than spending their budget on creature comforts, it impressed me that they were reserving it for research and technology.
Despite the Spartan feel, Tartan Racing’s headquarters seemed glamorous to me. I was a top executive in a major corporation, managing a budget in the billions of dollars at GM, but a part of me envied these young men. They weren’t the sort of people who crunched numbers in an ivory tower. They were “salt of the earth, roll in the mud, get your fingers dirty” engineers—who just happened to believe they could change the world. Shielded from the university bureaucracy by Whittaker, unencumbered by the sort of corporate red tape required by General Motors, these guys were getting things done that GM never could.
I watched as Whittaker, Urmson and the Tartan Racing team tested Boss in the sort of situations that had been modeled using computer simulations only weeks before. Boss’s first challenge at the demo I witnessed was a three-way intersection. The rules the programmers had coded into Boss’s digital brain guided the robot to yield to other vehicles that had arrived before it, and when it was Boss’s turn, the robot rolled through without a problem—an enormous relief to the team. As an additional feature of complexity, the team incorporated bad drivers into the equation. In one such test conducted at Robot City, Boss arrived at an intersection after a white American-model sedan, followed by a third vehicle—the second Humvee donated by AM General before the second Grand Challenge. The white sedan went first. Boss inched forward—and then the Humvee sped through the intersection, completely out of turn. Rather than rolling into the Humvee, Boss paused—exactly the appropriate action.
I was so excited I asked to go for a ride in Boss, which earned me some looks. The team didn’t seem certain it was a good idea. But I insisted, and moments later, I was squeezing myself into the only bit of interior cockpit space not occupied by computer equipment or batteries. Soon I grasped how different this robot was from the sort of vehicles my team designed at GM. Boss accelerated toward a stop sign and braked at the last minute. It careened around a turn and spun its tires over potholes and rocks without slowing. Every motion happened with a jerk or a slam. Just a minute or two into the ride I felt carsick for the first time in years. Now I understood their hesitation. Boss wasn’t designed for human occupancy, Whittaker explained soon after my ride. Rather, it had been designed for the express purpose of winning DARPA’s third race. The robot had been programmed to accelerate aggressively, and brake hard when the situation warranted. The herky-jerky behavior made humans motion sick—but it also made Boss fast.
After an open house inviting the public to witness a demonstration of Boss’s autonomous driving, the Tartan team loaded Boss into a tractor trailer and drove it to Arizona, where my staff had arranged for the robot to be tested at the GM Proving Grounds at Mesa. The warmer weather and the wide-open space of GM’s test tracks allowed Boss to begin learning how to behave in parking lots—which DARPA had said would be a key part of the Urban Challenge. The team modeled left-hand turns into traffic, a difficult proposition for many human drivers. And when the temperatures warmed up in the spring, Boss returned to Pittsburgh to prepare for a visit from DARPA, during which the Urban Challenge’s program manager, Norm Whitaker (no relation to Red), would monitor the sentient Chevy Tahoe as the robot conducted a series of tests. How the robot performed would dictate whether Tartan Racing would be one of the teams that remained in the race when DARPA staged the next round of eliminations. The resulting shortlist would compete at the so-called semifinals, the national qualifying event at the end of October 2007.
During the site visit, Boss had to pass four different tests. These challenges were conducted on Robot City’s quarter-mile track. Norm Whitaker and a crowd of about a hundred, including media and sponsors, watched as Urmson and his team demonstrated that Boss’s emergency stop button could halt the vehicle at a moment’s notice. The robot had to pass through an intersection also used by other vehicles without a collision, which it did. Another challenge involved driving down a street and avoiding a parked car—no problem. “Boss behaved like a good beginning driver,” praised DARPA’s Norm Whitaker. “A real good job.”
At the beginning of August, DARPA director Tony Tether announced the names of the thirty-five finalists invited to compete at the national qualifying event in late October. Tartan Racing was among them, which Whittaker, Urmson and Salesky were expecting. What they didn’t expect was Tether telling a reporter that Boss was not considered to be among the top-five robots.
Consequently, the final months before the qualifiers saw Urmson and Saleskey conducting exhaustive testing of Boss to ferret out any bugs in the algorithms the programmers had constructed—to make Boss as safe and capable a driver as any human being. Some of this testing involved hiding an inflatable model of a car next to the road. Just as Boss was about to pass, a team member would shove the bubble car onto the pavement. Everyone watched to see whether Boss would react in time.
When the robot became expert at dealing with that situation, Urmson escalated the challenge, to using an actual car. One day in late summer the guys were testing in Arizona. Salesky was driving a rental car in front of Boss. “Okay,” Urmson, who was riding shotgun inside the robot, said to Bryan over a walkie-talkie. “I want to make sure the vehicle’s speed control is working. Slam the brakes.”
The first time Salesky slammed the brakes, Boss slowed to a stop. But it wasn’t abrupt enough for Urmson. “Slam them harder,” Urmson told his friend. Moments later Salesky swerved in front of the robot then jammed his foot down on the brake pedal. “Harder,” Urmson beseeched Salesky. “Make the tires skid.”
This was Red Whittaker’s modus operandi. It’s only in the edge cases, Whittaker liked to say, the test-to-failure cases, that you learn about the robot’s capabilities.
Salesky shrugged—and slammed on the brakes.
And Boss ran straight into the back of the rental car. Salesky climbed out and went around to look at the rear bumper. The back of the rental had crumpled like a spent piece of paper. The damage to Boss was worse, because many of its most sensitive instruments, including a pair of medium-range radar units, were bumper-mounted. “The collision blew up about ten grand worth of sensors,” Salesky recalls. Urmson and Salesky stood there, arms akimbo, regarding the damage and shaking their heads. “Why did we just do this?” Salesky asked.
“My fault,” Urmson said.
“No, I should have known not to do it that hard,” Salesky said. He laughs about it, a decade later. “We both knew it was completely unrealistic. The following distance was way too close. There’s moments when you’re testing for so many hours in the field and you’re not really thinking clearly—and that was one of those moments.”
Sometimes, during such testing, a representative from DARPA dropped in on Tartan Racing